Tagging Syllable Boundaries With Joint N-Gram Models
Helmut Schmid, Bernd Möbius, Julia Weidenkaff
Institute of Natural Language Processing (IMS), University of Stuttgart, Germany
{schmid,moebius,weidenja}@ims.uni-stuttgart.de
Abstract
This paper presents a statistical method for the segmentation of
words into syllables which is based on a joint n-gram model.
Our system assigns syllable boundaries to phonetically transcribed words. The syllabification task was formulated as a
tagging task. The syllable tagger was trained on syllableannotated phone sequences. In an evaluation using ten-fold
cross-validation, the system correctly predicted the syllabification of German words with an accuracy by word of 99.85%,
which clearly exceeds results previously reported in the literature. The best performance was observed for a context size of
five preceding phones. A detailed qualitative error analysis suggests that a further reduction of the error rate by up to 90% is
possible by eliminating inconsistencies in the training database.
Index Terms: syllabification, joint n-gram models, German
1. Introduction
Automatic syllabification of words is an important task in a
number of natural language processing and speech technology
applications. Knowledge of syllable structure is crucial for assigning phone durations and F0 contours as well as selecting
appropriate units in concatenative speech synthesis [1]. It has
also been shown to improve word modeling in automatic speech
recognition [2]. Early approaches to automatic syllabification
were knowledge-based and built on the maximum onset principle [3] or the sonority hierarchy [4, 5]. Recently, work on
data-driven syllabification has been presented [6, 7, 12].
Syllabification can be performed either on the orthographic
representation of words or on the phonetic symbol string (phonetic transcription) of words, depending on the particular task
at hand. The former design may be useful under the assumption
that knowledge of syllable structure can improve pronunciation
[7]. The latter design is sometimes applied in speech synthesis for languages, such as German, in which pronunciation is
known to be sensitive to morphological structure and syllabification can be performed more reliably based on the phonetic
transcription than on orthography [8].
The syllable structure of German is complex. German
phonotactics allows consonant clusters in the onset and coda of
syllables. The maximum number of consonants is 3 in the onset
(e.g., [StR]) and 5 in the coda (e.g., [mpfst]).1 Thus, a maximum
number of 8 consecutive consonants may occur across syllable
boundaries. This complexity of onset and coda structures poses
problems for a syllabification algorithm because multiple alternative syllable boundary locations are usually possible in polysyllabic words.
The automatic syllabification method presented in this paper aims at assigning syllable boundaries to phonetically transcribed words. This task is formulated as a tagging task: each
1 SAMPA
notation is used for phonetic transcriptions in this paper.
phone symbol in the transcription of a word is annotated as
either preceding a syllable boundary or not preceding a syllable boundary. Our system learns the probabilities of syllable
boundaries from annotated corpora and predicts the locations
of syllable boundaries in previously unseen, unsyllabified, transcriptions of words.
In a series of experiments, the performance of the syllable
tagger was evaluated on the German part of the CELEX lexical
database [9]. The syllable tagger was found to predict correctly
the syllable boundaries in words from test data held out from
the training set with an accuracy by word of 99.85%.
It is important to note that the syllabification information
given by the CELEX database is used in our experiments both
as an evidence base from which the syllabification of the test
data is inferred, and as a reference syllabification for evaluation. This is an established procedure in the absence of a gold
standard [7]. Therefore, the performance of the syllable tagger
is a function of, among other factors, the quality of the evidence
base. This is not a trivial statement, as the qualitative error analysis in section 3.1 will illustrate.
2. Syllable tagger
We have chosen a tagging approach to syllabification: our syllabification program annotates each phone symbol in the transcription of a word either with a ‘B’ tag (indicating a syllable
boundary after the phone) or an ‘N’ tag (no syllable boundary). For instance, the correct tagging of the phone sequence
[pake:t@] (Pakete ‘packages, parcels’) is ‘p/N a/B k/N e:/B t/N
@/N’ [pa . ke: . t@].
2.1. Statistical model
Our syllabifier chooses the most likely tag sequence b̂n
1 =
b̂1 , b̂2 , ..., b̂n for the given phone sequence pn
1 . In other words,
it chooses the tag sequence which maximizes the conditional
n
probability P (bn
1 |p1 ) according to equation 1. The prior probability P (pn
)
of
the
phone sequence in equation 2 is indepen1
dent of the tag sequence bn
1 . Therefore, it has no influence on the
ranking of the different tag sequences and can be ignored, which
leads to equation 3. Equation 4 is obtained by decomposing the
probability of the tagged phone sequence into a product of conditional probabilities. After introducing a Markov assumption
(i.e. we assume that each tag–phone pair only depends on the k
preceding tags and phones, and that the probabilities are timeinvariant), equation 4 simplifies to equation 5.
b̂n
1
=
n
arg max
P (bn
1 |p1 )
n
(1)
=
n
n
arg max
P (bn
1 , p1 )/P (p1 )
bn
1
(2)
=
n
arg max
P (bn
1 , p1 )
bn
1
(3)
b1
=
=
arg max
n
b1
arg max
n
b1
n
Y
i−1
P (bi , pi |bi−1
1 , p1 )
(4)
i=1
n+1
Y
i−1
P (bi , pi |bi−1
i−k , pi−k )
(5)
i=1
In order to make sure that all conditional probabilities in
equation 5 are well-defined, we set pi = # and bi = N for i <
1. (The dummy phone symbol # is used to indicate the word
boundary.) Furthermore, we define pk+1 = # and bk+1 = N
to mark the end of the word, and we multiply the conditional
probabilities from position 1 to n + 1 rather than n.
Without the last refinement, the erroneous syllabification
[za:k . t] (sagt ‘says’) would be probable, because it follows
n
from equation 5 that the probability of any sequence hbn
1 , p1 i
must be at least as high as the sum of the probabilities of
n
all sequences which start with hbn
In our example,
1 , p1 i.
[za:k . t] (sagt) must be more probable than [za:k . t@] (sagte),
[za:k . t@st], (sagtest), [za:k . t@n] (sagten), and [za:k . t@t]
(sagtet) together, although the last syllable [t] is not a valid German syllable.
Our syllable tagger uses the Viterbi algorithm [10] to efficiently compute the most probable tag sequence according to
equation 5.
2.2. Parameter smoothing
The model parameters P (b, t|C) with C = bk1 , pk1 need to be
smoothed in order to avoid problems with zero probabilities.
We used the following simple backoff strategy which smoothes
frequencies by adding the backoff probability which is multiplied by some predefined weight parameter Θ. The smoothed
frequencies are then divided by the context frequency plus Θ to
obtain the probability estimates.
P (b, t|C) =
F (C, b, t) + ΘP (b, t|C 0 )
F (C) + Θ
The example in Figure 1 shows how the context C is generalized to the backoff context C 0 . The initial phone ([E]) of the
context, consisting of the phone sequence [Efl] and the pertinent
syllable boundary tags, is first generalized to a flag indicating
whether the phone was a vowel (+) or a consonant (–), i.e., it
is replaced by ‘+’ in the present example. In the second generalization step, this flag is deleted together with the respective
syllable boundary tag (‘N’), leaving a shorter context phone–tag
sequence. The same procedure is iteratively applied, if necessary, to the remaining context, which can be maximally reduced
to zero.
E/N f/B l/N
+/N f/B l/N
f/B l/N
-/B l/N
l/N
-/N
<zero>
Figure 1: Illustration of the backoff strategy for parameter
smoothing; see text for details.
The exact value of the smoothing parameter Θ had little
influence on the syllabification accuracy.
2.3. Syllable filter
Initial tests indicated that the syllable tagger described above
sometimes produced syllables containing two vowels, as in
[trEntSko:ts] (Trenchcoats), or without a vowel, like the last syllable in [rau . bau . ts] (Raubauz ‘brute’). However, vowel-less
syllables or syllables comprising two or more vowels violate
general constraints on the syllable structure of German, which
can be stated by the regular expression C*VC*.2
We modified the tagger in order to make sure that the tagger
only produces syllables with exactly one vowel. Each state of
the new tagger corresponds to a k-tuple of tag–phone pairs (as
before) plus an additional flag indicating whether or not the current partial syllable already contains a vowel or not. This flag
is cleared in the start state. If a vowel is encountered, the flag
is set (i.e., the flag of a state which emits a vowel must be set).
After a syllable boundary, the flag is cleared again.3 If a vowel
is immediately followed by a syllable boundary, the flag is also
cleared. In all other cases the vowel flag remains unchanged.4
Syllables with zero or two vowels are excluded by the following additional restrictions: (i) If the vowel flag is set, the
next phone cannot be a vowel (eliminating syllables with two
vowels).5 (ii) If the vowel flag is not set, the next “phone” cannot be the word boundary symbol ‘#’ (eliminating final syllables without a vowel).6 (iii) If the vowel flag is not set and the
next phone is a consonant, the next boundary tag must be ‘N’
(eliminating non-final syllables without a vowel).7
This modification also increased the speed of the syllable
tagger due to the smaller search space.
3. Evaluation
We evaluated the syllable tagger on the German part of the
CELEX lexical database [9], which contains 309,738 different
single-wordform entries.8 98.2% of the wordforms had more
than one syllable. The average number of syllables was 3.6.
The words in the database were split into 10 subcorpora according to two different designs. In test A, the database entries
were randomly assigned to the 10 subcorpora. In test B, all
wordforms pertaining to the same lemma were assigned to the
same subcorpus; otherwise, the assignment of database entries
to subcorpora was again random. The design of test B was motivated by the possibility that for a given inflected wordform in
the test set, the syllabifier has seen other, very similar, wordforms of the same lemma in the training set. In test C, we investigated whether stress information improves the performance of
the tagger. We used the same division into 10 subcorpora as in
test B and we encoded the stress with a pseudo phone [’] which
was inserted in front of the vowel of the stressed syllable. The
word sechstausend (six thousand) with two stress accents was
2 Note that in actual pronunciation, sonorants can be syllabic in German, for instance as a consequence of schwa deletion in unstressed syllables. However, the CELEX database, which provides the training and
test material for our syllabifier, represents canonical transcriptions.
3 If the last boundary tag is ‘B’, the vowel flag must be cleared.
4 The vowel flag of the next state is identical to the vowel flag of the
preceding state.
5 There is no transition from a state whose vowel flag is set to a state
whose last phone is a vowel.
6 There is no transition from a state whose vowel flag is not set, to a
state whose last “phone” is the word boundary symbol.
7 There is no transition from a state whose vowel flag is not set, to a
state whose last “phone” is a consonant and whose last tag is ‘B’.
8 Entries for multi-word tokens were deleted and entries differing
neither in the phone sequence nor the syllable structure were merged.
CV
prec.
99.94%
recall
99.94%
f-score
99.94%
acc.
99.85%
Table 1: Syllable boundary tagging results for test design
A. Ten-fold cross-validation using optimal context size and
smoothing parameter values (k = 5, Θ = 10−2 ).
represented as [z’Ekst’auz@nt] in test C.
In a series of experiments, we used the subcorpora 1–9 for
training and subcorpus 10 for testing. The phone context size k
and the smoothing parameter Θ were systematically varied.
CV
prec.
99.32%
recall
99.32%
f-score
99.32%
acc.
98.26%
Table 2: Results for test B, wordforms of the same lemma in the
same subcorpus. Ten-fold cross-validation, k = 4, Θ = 10−4 .
CV
prec.
99.32%
recall
99.31%
f-score
99.31%
acc.
98.23%
Table 3: Results for test C using stress information. Ten-fold
cross-validation, k = 4, Θ = 10−5 .
3.1. Test A
Table 1 shows the best results for the syllable boundary prediction task obtained after ten-fold cross-validation on randomly
split subcorpora of CELEX. In terms of absolute numbers, this
means that out of the total of 309,738 words, only 465 words
are assigned a syllabification that differs from the reference syllabification. In the overwhelming majority of words, there is an
accurate match between the syllabification assigned by our syllable tagger and those given by the CELEX database. Precision9
and recall are identical because the one-vowel-per-syllable constraint prevents the tagger from inserting additional syllables
or merging two syllables into one. The number of syllables is
therefore always correct. All remaining errors resulted from
shifting a syllable boundary over neighboring consonants.
The results of this experiment indicate that a context size
of less than 4 preceding phones produces a suboptimal, yet still
very good, syllabification performance. On the other hand, relatively large contexts (5 or 6 phones) are fairly reliable, despite
the fact that data sparsity increases with context size. Evidently,
the backoff strategy described in section 2.2 effectively alleviates the sparse data problem. The results obtained with small
values of the smoothing parameter are close to optimal. The
probability mass assigned to unobserved events is actually vanishingly small. The best context size was 5, but a context size
of 4 or 6 is almost equally good.
We manually inspected the first 97 words for which the syllabification assigned by the syllabifier differed from the one
provided by the CELEX database. The following error types
were identified:
• 48 words were probably mistagged because of inconsistencies in the CELEX database.
• 1 error was caused by a phone error in CELEX.
• In 38 cases, the syllabification probably failed because
of missing glottal stop information.10
• 10 differences to the reference syllabification were definitely syllabification errors. Examples are the word
Genugtuungen ‘satisfactions’ [g@ . nu:k . tu: . U . N@n],
which was syllabified as [g@ . nu:k . tu: . UN . @n];
and the word verfälschten ‘falsified’ [fEr . fElS . t@n],
which was syllabified as [fEr . fEl . St@n]. 4 out of these
10 errors occurred with foreign words, such as Ragtime
[r{g . taIm], which was syllabified as [r{ . gtaIm]. This
9 Precision is the number of correctly predicted syllables devided by
the total number of predicted syllables. Recall is the number of correctly
predicted syllables divided by the total number of correct syllables. And
the f-score is the harmonic mean of precision and recall.
10 C ELEX transcriptions do not include glottal stops, because it is assumed that glottal stops can be assigned by post-lexical rules.
error was probably caused by a sparse data problem: The
vowel “{” only occurred in six lemmata.
From this error analysis, we conclude that a further reduction of the error rate by up to 90% is possible by eliminating inconsistencies in the CELEX data and adding information about
glottal stops.
3.2. Test B
Table 2 shows the results obtained for test design B in which
all wordforms pertaining to the same lemma were assigned to
the same subcorpus; otherwise, database entries were randomly
assigned to the 10 subcorpora. The best context size and the
best smoothing parameter value observed in an evaluation on
subcorpus 10 were k = 4 and Θ = 10−4 , respectively. Table 2
reports the results in terms of precision, recall, f-score, and syllabification accuracy by word for the ten-fold cross-validation
using these optimal parameter values. The slight drop in performance suggests that in test A the syllabifier capitalized indeed
on the random distribution of wordforms of the same lemma
across subcorpora.
3.3. Test C
In test C, the tagger had access to information about the stress
which was added as a pseudo phone in front of the vowel of
the stressed syllable. Otherwise, test C was identical to test B.
We tested this version using subcorpora 1–9 as training data and
subcorpus 10 as test data. For very small contexts (k = 2), the
stress information improved the tagger performance by about
0.3%, but for larger contexts (k > 2), the accuracy was up to
0.2% lower. Table 3 shows the results obtained with ten-fold
cross validation using the optimal parameters from the evaluation on subcorpus 10. These results are almost identical to the
results of test B. We conclude that the stress information is not
relevant for the syllabification task.
4. Discussion
Syllable boundary prediction in German has been extensively
studied by [11]. She experimented with probabilistic contextfree grammars and multivariate clustering models. Her best
system, evaluated on the CELEX data, achieved 96.88% word
accuracy. We evaluated our system also on her data (which was
easier than our data) and obtained a word accuracy of 99.98%.
The work most similar to our approach was done by [13]
and used a standard trigram POS tagger [14]. The boundary tag
set was more complex than our binary tag set, but the context
was limited to two preceding phones. The system was evalu-
ated on CELEX data and yielded a tagging accuracy of 98.34%.
This result is very close to the word accuracy of 98.32% of our
syllable tagger for the same context size of k = 2, but it is not
clear whether the percentage of correctly syllabified words was
reported or the percentage of correctly assigned boundary tags.
The latter task should be simpler than the correct syllabification
of complete words.
Accuracies higher than 98% are unusual in linguistic processing. Why does the syllable tagger perform so extremely
well? One possible reason is that German is a language with
rich inflectional paradigms and productive word formation processes. In any text corpus as well as in the wordform database
of CELEX, nominal, adjectival and verbal stems occur with different inflectional suffixes and as the bases for many different
derivations and compounds. So, for many words in the test
data, there was a word with the same base form, but a different inflection, derivation or composition, in the training data.
Similarly, for any inflectional or derivational affixes occurring
in the test data, there is usually a number of occurrences of base
words with the same affix in the training data. The syllable tagger evidently combines these information sources to deduce the
correct syllabification of the word.
This consideration was partly taken into account in the test
B part of the evaluation, in which all inflectional wordforms of a
given lemma were forced into the same subcorpus. This experimental design assured that during the ten-fold cross-validation
the syllable tagger was prevented from seeing wordforms in the
training set that are just inflectional variants of the test word
at hand. Indeed, the slight drop in performance observed as a
consequence of this design, relative to the results of the experiments on randomly split subcorpora, suggests that structural information obtained from closely related wordforms constitutes
a valuable source of information for the syllable tagger. A more
radical test design aiming at keeping all morphologically related
wordforms in the same bin, was impractical for the present set
of experiments. It is also debatable whether such a design is
really desirable: after all, exploiting structural properties and
parallelisms between morphologically related wordforms is at
the heart of the work on supervised learning of syllabification
as it is presented in this paper.
The size of the context window has a large influence on the
accuracy of the system. Our results indicate that a context size
of 5 preceding phones produces optimal syllabification results.
A context size of 4 phones is sufficiently informative to obtain nearly optimal results, and a larger context (6 phones) also
yields a nearly optimal performance despite an increased sparsity of training data. However, a context of less than 4 preceding
phones produces a suboptimal syllabification performance. The
optimal context size can be interpreted with respect to the syllable structure of German. The average syllable length in terms of
the number of phones per syllable is 3.73 in the German wordform database of CELEX. This means that for a context size of
4 phones, there is a fair chance of seeing all previous phones
of the current syllable. For a context size of 5, the probability is quite high that the last phone of the preceding syllable is
also seen, which entails that the preceding syllable boundary is
included in the context. Thus, the experimentally determined
optimal context size is compatible with an approximate coverage of complete syllables by the training data.
5. Conclusions
We presented a probabilistic approach to automatic syllabification which assigns syllable boundaries to phonetically tran-
scribed words. The syllabification task was formulated as a
tagging task. The syllable boundary tagger is based on a joint
n-gram model. It was trained on syllable-annotated phone sequences drawn from the German CELEX wordform database.
Using ten-fold cross-validation, the syllable tagger correctly
predicted the syllabification of words with an accuracy by word
of 99.85%. The joint n-gram model requires a context of at
least four preceding phones to achieve a close to optimal performance; the best performance was observed for a context size of
five preceding phones. The syllabification performance clearly
exceeds results previously reported in the literature.
6. References
[1] Sproat, R., Ed., Multilingual Text-to-Speech Synthesis:
The Bell Labs Approach, Kluwer, 1998.
[2] Greenberg, S., “Speaking in shorthand—a syllable-centric
perspective for understanding pronunciation variation”,
Speech Comm., 29(2–4):159–176, 1999.
[3] Kahn, D., “Syllable-based generalizations in English
phonology”, Ph.D. diss., MIT, 1976.
[4] Clements, G.N. “The role of the sonority cycle in core
syllabification”, in Papers in Laboratory Phonology I: Between the Grammar and Physics of Speech, Kingston, J.,
Beckman, M.E., Eds., Cambridge UP, 283–333, 1990.
[5] Blevins, J., “The syllable in phonological theory”, in The
Handbook of Phonological Theory, Goldsmith, J.A., Ed.,
Blackwell, 206–244, 1995.
[6] Daelemans, W. and van den Bosch, A., “Generalisation
performance of back-propagation learning on a syllabification task”, in TWLT3: Connectionism and Nat. Lang.
Process., Drossaers, M. and Nijholt, A., Eds., Twente
Univ. Enschede, 27–37, 1992.
[7] Marchand, Y. and Damper, R.I., “Can syllabification improve pronunciation by analogy of English?”, J. Nat.
Lang. Engin., 1(1):1–25, 2006.
[8] Möbius, B., “The Bell Labs German text-to-speech system”, Comp. Speech Lang., 13:319–358, 1999.
[9] Baayen, R., Piepenbrock, R. and Gulikers, L., “The
CELEX lexical database (release 2)”, Linguistic Data
Consortium, Univ. Pennsylvania, 1995.
[10] Manning, C.D. and Schütze, H., Foundations of Statistical
Natural Language Processing, MIT Press, 1999.
[11] Müller, K., “Automatic detection of syllable boundaries
combining the advantages of treebank and bracketed corpora training”, Proc. 39th Ann. Meet. ACL, Toulouse,
402–409, 2001.
[12] Kiraz, G.A. and Möbius, B., “Multilingual syllabification
using weighted finite-state transducers”, Proc. Third Internat. Workshop on Speech Synthesis, Jenolan Caves, Australia, 71–76, 1998.
[13] Krenn, B., “Tagging syllables”, Proc. Eurospeech-97,
Rhodes, Greece, 991–994, 1997.
[14] Brants, T., “TnT - a statistical part-of-speech tagger”,
Proc. Sixth Appl. Nat. Lang. Process. Conf. ANLP-2000,
Seattle, WA, 2000.